IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO.
2, FEBRUARY 2002
Delayed-Pilot Sampling for Mixture Kalman Filter
With Application in Fading Channels
Xiaodong Wang, Member, IEEE, Rong Chen, Member, IEEE, and Dong Guo
Abstract—Sequential Monte Carlo (SMC) methods are pow-
erful techniques for online filtering of nonlinear and non-Gaussian
dynamic systems. Typically dynamic systems exhibit strong
memory effects, i.e., future observations can reveal substantial
information about the current state. Recently, the delayed-sample
sampling method has been proposed in the context of mixture
Kalman filter (MKF), which makes use of future observations in
generating samples of the current state. Although this method is
highly effective in producing accurate filtering results, its compu-
tational complexity is exponential in terms of the delay, due to the
need to marginalize the future states. In this paper, we address this
difficulty by developing two new sampling schemes for delayed
estimation, namely, delayed-pilot sampling and hybrid-pilot sam-
pling. The basic idea of delayed-pilot sampling is that instead of
exploring the entire space of future states, we generate a number
of random pilot streams, each of which indicates what would
happen in the future if the current state takes a particular value.
The sampling distribution of the current state is then determined
by the incremental importance weight associated with each pilot
stream. The delayed-pilot sampling can be used in conjunction
with the delayed-sample method, resulting in a hybrid scheme.
This new sampling technique is then applied to solve the problem
of adaptive detection and decoding in flat-fading communication
channels. Simulation results are provided to demonstrate the
performance of the new low-complexity sampling techniques
for delayed estimation and to compare with the delayed-sample
method.
Index Terms—Adaptive decoding, delayed-pilot, delayed-
sample, fading channel, mixture Kalman filter, sequential Monte
Carlo.
I. INTRODUCTION
T HE SEQUENTIAL Monte Carlo (SMC) methods recently
emerged in the fields of statistics and engineering have
shown a great promise in solving a wide range of nonlinear fil-
tering/prediction problems associated with many highly com-
plex dynamic systems [1], [3], [8], [12], [15], [18], [21]–[23],
[28]. SMC can be loosely defined as a family of methodolo-
gies that use Monte Carlo simulation to solve on-line estimation
problems in dynamic systems. By recursively generating Monte
Manuscript received January 31, 2001; revised October 3, 2001. This work
was supported in part by the National Science Foundation under Grants CCR-
9980599, DMS-9982846, DMS-0073651, and DMS-0073601. The associate
editor coordinating the review of this paper and approving it for publication
was Dr. Petar M. Djurić.
X. Wang was with the Department of Electrical Engineering, Texas A&M
University, College Station, TX 77843 USA. He is now with Department of
Electrical Engineering, Columbia University, New York, NY 10027 USA.
R. Chen is with the Department of Information and Decision Science, Uni-
versity of Illinois at Chicago, Chicago, IL 60607 USA.
D. Guo is with the Department of Electrical Engineering, Texas A&M Uni-
versity, College Station, TX 77843 USA.
Publisher Item Identifier S 1053-587X(02)00563-9.
1053–587X/02$17.00 © 2002 IEEE
241
Carlo samples of the state variables or some other latency vari-
ables, these methods can flexibly adapt to the dynamics of the
underlying stochastic systems. A complete theoretical frame-
work for the SMC has appeared in [23]. Many successful appli-
cations of SMC in diverse areas of science and engineering can
be found in the recent books [9], [20].
Recently, we have applied the SMC technique, and in
particular, the mixture Kalman filter (MKF) for conditional
dynamic linear models [3], to the problem of adaptive detection
and decoding in flat-fading channels [4] in the presence of
Gaussian or non-Gaussian ambient noise. This problem is
of fundamental importance in communication theory, and it
is known that the optimal solution has a prohibitively high
computational complexity [13], [25], [26]. Various suboptimal
approaches have been proposed in the literature to bear on this
problem [2], [5]–[7], [10], [11], [14], [16], [17], [19], [24],
[27], [29], [30]. Our approach in [4] is based on a Bayesian
formulation of the problem of on-line detection/decoding in
flat-fading channels and the SMC methodology. The basic idea
is to sequentially impute multiple samples of the transmitted
symbols based on the current observation. Associated with
each sequence of imputed symbols is its important weight.
The imputed symbol sequences, together with their importance
weights, are used to compute the Bayesian estimates of the
transmitted information. It is shown in [4] through simulations
that the performance of the SMC-based receivers can be
remarkably close to the so-called genie-aided bound in fading
channels for both uncoded and coded systems without the use
of any training/pilot symbols or decision feedback.
Dynamic systems often possess strong memory, i.e., future
observations can reveal substantial information about the cur-
rent state. For example, in a flat-fading communication channel,
the channel fading coefficients are modeled as a correlated com-
plex Gaussian process. Even if the transmitted symbols are inde-
pendent, since the observations in the near future contain future
channel states that are highly correlated with the current channel
state, it is beneficial to make use of these future observations in
detecting the current symbol. Moreover, when channel coding
is employed, the transmitted symbols are also highly correlated,
and it becomes crucial to exploit future observations to extract
useful information contained in both future symbols and future
channel states. However, an SMC method usually does not go
back to regenerate past samples in view of new observations,
although the past estimation can be adjusted by using the new
importance weights. To overcome this difficulty, in [4], we have
developed the delayed-sample method, which makes use of fu-
ture observations in generating samples of the current state. It is
seen there that this method is especially effective in improving
242 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

the receiver performance when the signals are channel-coded In particular, if is sampled from a trial distribution that
and/or when the channel ambient noise is non-Gaussian. has the same support as , and let , then
However, the computational complexity of the de- is properly weighted with respect to .
layed-sample method developed in [4] is quite high, due Consider the following conditional dynamic linear model
to the need of marginalizing out the future states. For ex- (CDLM) of the form
ample, in a Gaussian noise flat-fading channel, for a -step
delayed-sample method, the algorithmic complexity is (1)
(where is the finite set of signal constellations, (2)
and is the number of Monte Carlo samples) for an uncoded
where and are the state and ob-
system, and it is for a coded system with a rate
servation noise, respectively, and is a sequence of random
convolutional code. The complexity is even much higher
indicator variables that may form a Markov chain but are inde-
if the ambient noise is non-Gaussian. Hence, it becomes im-
pendent of and and the past and , . The matrices
portant to develop computationally efficient delayed estimation
, , , and are known given . The indicator
techniques.
takes values from a finite set .
In this paper, new low-complexity techniques are developed The MKF in [3] is a novel sequential Monte Carlo method
for delayed estimation under the MKF framework. In order to for online filtering and prediction of CDLMs; it exploits
draw a new sample of the current state, instead of marginal- the conditional Gaussian property and utilizes a marginal-
izing the future states of steps, we generate random pilot
ization operation to improve the algorithmic efficiency. Let
streams of these future states. Each pilot stream is obtained by and . By
starting with one of the possible value and propagating recursively generating a set of properly weighted random
steps of SMC. The incremental importance weight of this
samples to represent , the MKF
pilot stream is then given to the value and with probability
approximates the target distribution by a random
proportional to which a sample of the current state is drawn. Fi-
mixture of Gaussian distributions
nally, to correct the bias introduced by the pilot approximation,
the usual incremental weight is multiplied by the inverse of the
pilot weight. We term this method of delayed estimation the de-
layed-pilot method. Moreover, such an idea can be combined
with the delayed-sample method developed in [4] to yield a de- where
layed-hybrid method. That is, in order to achieve a -step
and
delayed estimation, we implement the combination of a -step
delayed-sample sampling and a -step delayed-pilot sampling. are obtained with a Kalman filter on the system (1), (2) for the
The delayed-pilot method and the delayed-hybrid method are given indicator trajectory . Denote
then applied to the problem of adaptive detection and decoding
in flat-fading channels treated in [4].
The rest of this paper is organized as follows. In Section II,
we briefly summarize the MKF algorithm, as well as the de- Thus, a key step in the MKF is the production at time of
layed-weight method and delayed-sample method for MKF. In the weighted samples of the indicators and the Kalman filter
Section III, we develop the delayed-pilot method and the de- outputs , based on the set of samples
layed-hybrid method for MKF. In Section IV, we treat the prob- at the previous time [3], [4].
lems of adaptive detection and adaptive decoding in flat-fading The model (1) and (2) is often observed to be highly cor-
channels using the proposed new sampling methods. Simulation related. As a result, future observations often contain informa-
results are provided in Section V. Finally, Section VI contains tion about the current state. Hence, a delayed estimate is usually
the conclusion. more accurate than the concurrent estimate. In delayed estima-
tion, instead of making inference on or instantaneously
with the posterior distribution or , we delay
II. BACKGROUND this inference to a later time , with the distribution
First, we define an important concept that is used throughout and , respectively. In [4], two delayed
the paper. A set of random samples and the corresponding estimation techniques are proposed, namely, the delayed-weight
weights is said to be properly weighted with method and the delayed-sample method.
respect to distribution if, for any measurable function , The idea behind the simple delayed-weight method is as fol-
we have lows. If the set is properly weighted with
respect to , then when we focus our attention on
at time , we have that is properly
weighted with respect to . Then, any inference about
the indicator , can be approximated by
as
(3)
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER 243

with
(4)

Since the weights contain information about the fu-

ture observations , the estimate in (3) is usu-
ally more accurate than the concurrent estimate. This approach
only requires slight additional memory and no additional com-
putation. The interference about the state variable can be sim-
ilarly made at time .
In the delayed-sample method, both the samples and
the weights are generated based on the
observations , hence making the target
distribution at time . The procedure provides better
Monte Carlo samples since it utilizes the future observations
in generating the current samples of ,
but the algorithm is also more demanding both analytically
and computationally because of the need of marginalizing out
. At any time , the delayed-sample algorithm
involves growing and trimming a tree of depth . That is, we
grow the tree by adding leaves at the top of tree, then
keep one the main branches from the root corresponding to
the selected , and then removing the rest of the branches.
Fig. 1 illustrates the delayed-sample sampling operation, with
Fig. 1. Illustration of the delayed-sample sampling.
and .

operation, with and . The algorithm is sum-

III. DELAYED-PILOT SAMPLING IN MIXTURE KALMAN FILTER marized as follows.
A. Simple Delayed-Pilot Sampling Algorithm 1 [Delayed-Pilot MKF]: Suppose at time
, a set of properly weighted samples
The delayed-sample MKF algorithm can achieve significant
with respect to
improvement in performance compared with the concurrent
are available. Then, at time as the new data
MKF algorithm, as demonstrated in [4]. However, such im-
become available, the following steps are implemented to
provement comes at the expense of significant increase in
update each weighted sample.
computational complexity since it requires full exploration
For :
of the space of future states . On the other
hand, the delay-weight method utilizes the future observation • For each , run a one-step Kalman filer update as-
in only through the weight adjustments. The suming to obtain
sample of generated is still based only on the observation
up to time .
Because of the desire of utilizing future data and because of (5)
the high computational complexity associated with a full explo-
ration of the space of the future states, in this paper, we de- and compute
velop a low-complexity algorithm that partially explores (but
hopefully the most important part of) the space of future states
and utilizes this information to generate better (6)
samples of . Specifically, since takes possible discrete
values in , we generate “pilots,” each starting with one • For each , do the following:
of the possible values and propagating to using the For , repeat
concurrent MKF steps. Then, the incremental weight (from to — Let
) of the pilot starting with reveals what would
happen in the future ( to ) if is chosen to be ,
given . Hence, these incremental weights can For each , run a one-step Kalman filer update
be used as the sampling distribution in selecting . This sam- assuming to obtain
pling distribution helps us to generate better samples for the
current state, even though we did not fully explore the entire
space of future states. We call this algorithm the delayed-pilot
sampling method. Fig. 2 illustrates the delayed-pilot sampling (7)
244 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

Using

(12)

as the trial sampling distribution, draw a sample . Ap-

pend to to obtain . If is drawn,
then set .
• Compute the importance weight. If , then

(13)

• Do resampling if the effective sample size is below certain

threshold, as discussed later in Remark 4.
Fig. 2. Illustration of the delayed-pilot sampling. Remark 1 (Multiple Pilots): In the above algorithm, we can
also use multiple pilots for each group. That is, we can generate
groups of pilots, each with members. Each group starts
— For each , compute the sampling density with one of the possible values in . The sum of the incremental
weights of each group (propagated from to ) is then used
as the sampling probability for . This is particularly useful
when is large.
Remark 2 (Inference): The above algorithm generates, at
time , a weighted sample that is properly
(8)
weighted with respect to . This can be easily seen
from
Draw a sample according to the above sampling
density. If is drawn, then set .
— Compute the incremental importance weight

(9)

(14)

where the numerate of (14) is the target distribution, and the

denominator is the sampling distribution of according to
the algorithm.
(10) Unfortunately, is not properly weighted with re-
spect to . For the latter, one would have to ex-
plore the entire future space, as in the delayed-sample algorithm.
• For each , compute
Hence, the weighted average

(11)
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER 245

is an estimate of but not of . Remark 3 (Combination With Delayed-Weight): It is desir-

Hence, it seems that our purpose of using future data to help able to combine the delayed-pilot method with the delayed-
making inference is not served at all. Although the sample weight method. Specifically, suppose at time we obtain a
is closer to the distribution , the weight adjusts sample using the delayed-pilot method. Then, with
them backward to , hence giving smaller weights an additional delay , we can use the auxiliary weight at
to those close to but away from and time to make inference on , i.e.,
giving larger weight to those close to but away from
.
Fortunately, there is indeed an easy way that allows us to
make better inferences on at time . Suppose .
Then, we calculate the following auxiliary weight: with
(18)
(15)

This is the original weight multiplied by the incremental This is because is properly weighted with
weight of the sampled pilot propagated from to . respect to .
Let Alternatively, if one does not maintain two sets of weights
and is willing to wait additional ( ) steps, then the
weight generated at time from the above algo-
rithm can be used for inference. At that time, is a
be the selected pilot. Then, is properly weighted
properly weighted sample with respect to . Then,
with respect to . This can be easily seen from the
following: is properly weighted with respect to .
Since , we still take full advantage of the delay .
Remark 4 (Resampling): Resampling is an important step in
SMC. It rejuvenates the sampler when the effective sample size
becomes small, hence providing more efficient samples for the
future states. For more details, see [23]. Specifically, suppose at
time we have a set of random samples prop-
erly weighted with respect to . By treating the samples
as a discrete approximation of the posterior distribution, we can
generate another discrete representation as follows.
• Select a set of positive numbers , .
• For
— Randomly select from the set with
probability proportional to .
— If , then set the new weight associated
with this sample to be .
(16) • Return the new representation .
where the numerator of (16) is the target distribution, and the de- The new weight is also properly weighted with respect to
nominator is the sampling distribution of , the th stream .
up to time , and the selected th pilot path. The choice of the resampling probability directly affects
Hence, is properly weighted with respect to the efficiency of the algorithm. As in [23], one choice is
. Therefore, we have , which makes the resulting resampled set have
equal weights. In the delay-pilot algorithm, the auxiliary weight
is a better choice. As shown above, gives small
weights to samples that are close to the future distribution
with but away from the current distribution .
Using as the resampling probability is shortsighted. It tends
(17) to remove the samples that are less important at time but may
be important in the future (e.g., at time ). On the other
Note that no additional computation is needed to obtain the aux- hand, the auxiliary weight is based on the future distribution.
iliary weights; hence, it is recommended to keep two sets of This resampling procedure does not reduce the variance of the
weights in the propagation. However, the auxiliary weight are current weight distribution as much as resampling according
computed at every time step and cannot be used for further prop- to . However, since it keeps better samples (in the eyes of
agation (i.e., to replace ). It is solely used for inference of future), the weight distribution at latter time will be better.
at time . Note that with , the new weight is proportional to
246 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

, where the inverse of the incremental weight of

the sampled pilot is propagated from to .
When the weight distribution becomes very skewed, resample
according to the original weights may result in severe loss of
diversity of the sample. In this situation, it is recommended to
use , or , where
can vary according to the coefficient of variation of . For the
delayed pilot algorithm, seems to be a good choice.
Remark 5 (Complexity): Note that the computation required
for the above simple delayed-pilot sampling is the same as that
for the one-step delayed-sample method. Hence, the delayed-
pilot MKF algorithm is recommenced only for . At each
time, this algorithm requires one-step
Kalman filter updates to be calculated, whereas the delayed-
sample MKF requires Kalman updates at each time.

B. Delayed-Hybrid Sampling
The delayed-hybrid sampling method is a combination of the
delayed-sample method and the delayed-pilot method. Here, we
fully explore the state space of the immediate future and gen-
erate pilots to explore the space beyond that.
Let be the total delay time. A delayed-hybrid sam-
Fig. 3. Illustration of the delayed-hybrid sampling.
pling method generates pilots, where each takes one of
the possible paths of and makes addi-
tional concurrent MKF steps to time . Then, the For :
probability of selecting is proportional to the sum of the • For each , and for each ,
incremental importance weights of all pilots, starting with perform a one-step update on the corresponding Kalman
. Since we have explored the entire space of , filter , i.e.,
the weight calculation will be based on the target distribution
. Fig. 3 illustrates the delayed-hybrid sampling op-
eration, with , , and . (21)
Specifically, for each possible “future” (relative to time )
symbol sequence at time , i.e., Compute

we keep the value of a -step Kalman filter ,

where

(22)
(19)
• For every path of , do the following:
Denote For , repeat:
— Denote
(20)

The delayed-hybrid sampling MKF algorithm recursively (23)

propagates the samples properly weighted for
to those for and is summarized as follows. For each , run a one-step Kalman filer update
Algorithm 2 [Delayed-Hybrid MKF]: Suppose at time assuming to obtain
, a set of properly weighted samples
with respect to
are available. Then, at time , as the new data
become available, the following steps are imple-
mented to update each weighted sample. (24)
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER 247

— For each , compute the sampling density

(25)
(29)
Draw a sample according to the above • Do resampling if the effective sample size is below certain
sampling density. If , set threshold, as discussed in Remark 4 (Section III-B).
Note that at each time the above delayed-hybrid MKF
algorithm involves calculating
one-step Kalman filter updates, whereas the full delayed-sample
— Compute the incremental importance weight MKF algorithm requires Kalman updates at each
time.
Similar to the proof of (14), it can be easily shown that the
above algorithm generates, at time , a weighted
(26) sample , which is properly weighted with respect
to but not to .
Remark 6 (Auxilary Weight): Similar to the delay-pilot algo-
• For each , compute
rithm, we also need to resort to the auxiliary weight to make
better inferences on at time . However, it is
slightly more complicated for the delay-hybrid algorithm.
Suppose . Define the concurrent weight

(27)

Using

(28) (30)
where is the sampling probability (28). Then,
is properly weighted with respect to , which is the con-
current posterior distribution.
as the trial sampling distribution, draw a sample . Ap- Let the th path in be , and , where
pend to to obtain . Based on this sample, is the pilot path started from in the
form based on the results from the previous step. th sample stream. Define for
• Compute the importance weight. Suppose and and
; then for .
For every possible value of , define the single path
weight as

(31)
248 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

By the same argument as in standard SMC [23], we replace

by .

IV. ADAPTIVE DETECTION/DECODING IN FADING CHANNELS

VIA DELAYED-PILOT MKF
(32)
A. System Description
We consider a communication system signaling through a
flat-fading channel with additive white Gaussian noise. The
complex data symbols take values from a finite alphabet
set , and each symbol is transmitted over a
(33) flat-fading channel whose input–output relationship is given by
(37)
where , , and are defined by (22), (26), and
(28), respectively. where , , , and are the received signal, the fading
Define the auxiliary weight as channel coefficient, the transmitted symbol, and the additive
Gaussian noise at time , respectively. The processes , ,
(34) and are assumed to be mutually independent.
As in [4], it is assumed that the channel-fading process is
Rayleigh. That is, the fading coefficients form a complex
We next show that is properly weighted with re-
Gaussian process that can be modeled by the output of a lowpass
spect to . First, for any measurable function
Butterworth filter of order driven by white Gaussian noise
on , define an extended function on such that
for all .
(38)
Let be the constant that makes (31) an equality. It is easily
seen is a constant for all . By the construction of , we
have where ; ;
and is a white complex Gaussian noise sequence with unit
variance and independent real and complex components. We can
rewrite (37) and (38) in the following state-space model form:
(39)
(40)
where

.. .. .. .. ..
. . . . .

and

(35)
Hence .. ..
. .

B. Adaptive Detection in Uncoded System

Denote and . For the
case of the uncoded system, the transmitted symbols are as-
sumed to be independent, i.e.,
(41)

(36) When no prior information about the symbols is available, the

symbols are assumed to take each possible value in with equal
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER 249

probability, i.e., for . We , with and , where

are interested in computing the symbol a posteriori probabilities is the stationary covariance of and is computed analytically
from (39). All importance weights are initialized as .
(42)
The initial are randomly generated from the set .
In [4], we have solved this problem based on the concurrent Suppose at time , a set of properly weighted sam-
MKF algorithm and the delayed-sample MKF algorithm. Here, ples with respect to
we consider solving the same problem using the simple de- and corresponding transmitted symbols are avail-
layed-pilot MKF algorithm and the delayed-hybrid MKF algo- able. Then, at time , as the new data become
rithm discussed in Section III. By the state-space mode (39) and available, the following steps are implemented to update each
(40), this is quite straightforward—we only need to make the weighted sample.
following substitution of variables in Algorithm 1 and Algo- For :
rithm 2 respectively, namely
• For each possible information bit vector
, compute the corresponding vector to obtain

and (46)
(43) Then, run steps of Kalman filter update to obtain

C. Adaptive Decoding in Coded System

We next consider the problem of adaptive decoding in a (47)
convolutionally coded system signaling through a flat-fading
channel using the delay-pilot sampling technique. and compute
Consider a binary rate convolutional encoder of overall
constraint length . Suppose the encoder starts with an
all-zero state at time . The input to the encoder at time is
a block of information bits ; the encoder (48)
output at time is a block of code bits .
For simplicity, here, we assume that BPSK modulation • For each , do the following:
is employed. Then, the transmitted symbols at time are For , repeat
, where , . — Denote
(That is, if , and if .) Since
is determined by , so is . Hence, we
can write and the corresponding transmitted symbols

(44)

for some function determined by the structure of the en- For each , compute the corre-
coder. We also assume the information bits are independent. sponding transmitted symbols , and run steps of
Let be the received signals at Kalman filter update to obtain
time , and let be the channel states
corresponding to and . Recall that .
In addition, denote ; ;
(49)
. The Monte Carlo samples recorded at time
are , where — For each , compute the sampling density

contains the mean and covariance matrix of the channel state (50)
vector conditioned on and . That is

(45) Draw a sample according to the above sampling

density. If is drawn, then set .
As before, given the information bit sequence , the corre- — Compute the increment importance weight
sponding is obtained by a Kalman filter.
The adaptive decoding algorithms based on the delayed-pilot
MKF is as follows.
Algorithm 3 [Delayed-Pilot Adaptive Decoder]: Initial- (51)
ization (For ): Each Kalman filter is initialized as
250 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

• For each , compute • For every path of do the following:

For , repeat:
— Denote
(52)

Using

(57)
(53)
and the corresponding transmitted symbols

as the trial sampling distribution, draw a sample . Ap-

pend to to obtain . If is drawn,
then set . (58)
• Compute the importance weight. If , then
For each , compute the cor-
responding transmitted symbols , and run steps
Kalman filer update to obtain

(54)

(59)
• Do resampling if the effective sample size is below certain
threshold. — For each , compute the sampling density
The adaptive decoding algorithms based on the delayed-hy-
brid MKF is as follows.
Algorithm 4 [Delayed-Hybrid Adaptive Decoder]: Initial-
ization: Same as in Algorithm 4.
Suppose that at time , a set of properly
weighted samples with respect to
and the corresponding transmitted signals (60)
are available; then, at time , as the
new data become available, the following steps are Draw a sample according to the above sam-
implemented to update each weighted sample. pling density. If , set
For :
• For each , and for each
, compute the corresponding transmitted sym-
bols . Perform steps update on the corresponding — Compute the increment importance weight
Kalman filter , i.e.,

(55) (61)

and compute
• For each , compute

(56) (62)
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER
Using
(63)
as the trial sampling distribution, draw a sample ac-
cording to the above sampling density. Append to
to obtain . Based on this sample, form
using the results from the first step.
• Compute the importance weight. Suppose and
; then
(64)
• Do resampling if the effective sample size is below certain
threshold.
V. SIMULATION RESULTS
In this section, we provide some computer simulation exam-
ples to demonstrate the performance of the adaptive receiver in
fading channels that employs the proposed delayed-pilot sam-
pling techniques. As in [4], the fading process is modeled by the
output of a Butterworth filter of order driven by a com-
plex white Gaussian noise process. The cut-off frequency of this
filter is 0.05, corresponding to a normalized Doppler frequency
(with respect to the symbol rate ) , which is a
251
Fig. 4. BER performance of the concurrent MKF receiver, the delayed-weight
MKF receiver with  =2, and the simple delayed-pilot MKF receiver with
1=2 =2 ; in an uncoded system.
fast fading scenario. Specifically, the fading coefficients
are modeled by the following ARMA(3,3) process:
(65)
where . The filter coefficients in (65) are chosen
such that . It is assumed that BPSK modulation is
employed, i.e., the transmitted symbols . In all
algorithms, 100 Monte Carlo samples are used, i.e., .
The performance of the adaptive receivers based on the con-
current MKF, the delayed-weight MKF, and the delayed-sample
MKF is studied in [4], which will serve as baseline for com-
parison here. Note that as illustrated in [4], for uncoded sys-
tems, the performance of the simple delayed-weight method is
already very close to the “genie-aided lower bound,”1 and the
delayed-sample method offers not much further improvement.
Hence, it can be inferred that in this case, both the delayed-pilot
method and the delayed-hybrid method will also perform close
to this lower bound, which is indeed confirmed by simulations,
as shown in Fig. 4, where it is seen that the performance of
the combined delayed-pilot method and delayed-weight method
with and offers a slightly better performance than
the delayed-weight method with a delay .
It is also shown in [4] that for coded systems, the de-
layed-weight method is no longer effective in exploiting the
strong memory of the system, and it is necessary to employ
the delayed-sample method for adaptive decoding, whose
performance steadily improves as the delay increases. In
what follows, we illustrate the performance of the delayed-pilot
1The genie-aided lower bound is obtained as follows. We assume that at each
time t, a genie provides the receiver with an observation of the modulation-free
channel coefficient corrupted by additive white Gaussian noise with the same
~ = +~
variance, i.e., y ~  N (0
 n , where n )
;  . The receiver employs a
^ = f j~ g
Kalman filter to track the fading process based on the information provided by
the genie, i.e., it computes  E  Y . The symbol is then demodulated
^ = sign( f ^ g)
according to s y .
252 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002

Fig. 5. BER performance of the simple pilot-delayed MKF receiver with Fig. 7. BER performance of the delayed-hybrid MKF receiver with different
different delays in a coded system. delays in a coded system.

Fig. 8. BER performance of the delayed-hybrid MKF receiver with a fixed

Fig. 6. BER performance of the receiver employing the combined
1=12345
delayed-pilot method (with ; ; ; ; ) and delayed-weight method
1+1 = 5
total delay ( ) in a coded system.
(with = 10 ) in a coded system.
effective in improving the receiver performance. In addition,
increased performance improvement is obtained as the length
method and that of the delayed-hybrid method in coded
of the pilot stream increases. Moreover, a further performance
systems. The information bits are encoded using a rate 1/2
enhancement is achieved by using the delayed-weight method
constraint length 5 convolutional code (with generators 23 and
in conjunction with the delayed-pilot method. Figs. 7 and
25 in octal notation). The receiver implements the adaptive
8 illustrate the performance of receiver employing the de-
decoding algorithms discussed in Section IV in combination
layed-hybrid method. Inferences are made using the auxiliary
with the delayed-weight method. That is, the information
weight at time . It is seen that
bits samples are drawn by using the delayed-pilot
this scheme offers further performance improvement over the
method with delay or the delayed-hybrid method with delay
delayed-pilot method (at the expense of increased computa-
, whereas the importance weights
tional complexity). Finally, the performance of the combined
are obtained after a further delay of .
delayed-hybrid method and the delayed-weight method is
Figs. 5 and 6 illustrate the performance of receiver em-
shown in Fig. 9.
ploying the simple delayed-pilot method with or without
combining with the delayed-weight method. In the same
figures, the performance of the concurrent MKF, as well as that VI. CONCLUSIONS
of the delayed-sample MKF with a delay , is shown. In this paper, we have developed a new sequential Monte
It is seen that the proposed delayed-pilot method is quite Carlo sampling method—the delayed-pilot sampling—under
WANG et al.: DELAYED-PILOT SAMPLING FOR MIXTURE KALMAN FILTER
Fig. 9. BER performance of the receiver employing the combined
delayed-hybrid method and delayed-weight method in a coded system.
the framework of mixture Kalman filter for delayed esti-
mation in conditional dynamic linear systems. Instead of
marginalizing the future states, as in the delayed-sample
method [4], this new scheme generates random pilot streams
using sequential important sampling, based on the future
observations. The incremental importance weights of these
pilot streams are then used to form the sampling density
of the current state as well as the importance weights as-
sociated with the samples of the current state. The compu-
tational complexity of the delayed-pilot MKF algorithms is
for a -step delay (where is the finite set
of possible states), whereas the corresponding complexity of
the delayed-sample MKF algorithm in [4] is . This
technique can also be used in conjunction with the delayed-
sample method, resulting in a hybrid sampling technique. For
a -step delayed-sample followed by a -step delayed-pilot,
the complexity of the corresponding delayed-hybrid algo-
rithm is , whereas the full delayed-sample
MKF algorithm has a complexity of for a
total delay of steps. Moreover, the performance of
both the delayed-pilot method and the delayed-hybrid method
can be further enhanced when employed in conjunction with
the delayed-weight method.
REFERENCES
[1] D. Avitzour, “A stochastic simulation Bayesian approach to multitarget
tracking,” Proc. Inst. Elect. Eng., Radar, Sonar, Navigat., vol. 142, pp.
41–44, 1995.
[2] J. K. Cavers, “An analysis of pilot symbol assisted modulation for
Rayleigh fading channels,” IEEE Trans. Veh. Technol., vol. 40, pp.
686–693, Nov. 1991.
[3] R. Chen and J. S. Liu, “Mixture Kalman filters,” J. R. Statist. Soc. B, vol.
62, no. 3, pp. 493–509, 2000.
[4] R. Chen, X. Wang, and J. S. Liu, “Adaptive joint detection and decoding
in flat-fading channels via mixture Kalman filtering,” IEEE Trans. In-
form. Theory, vol. 46, pp. 2079–2094, Sept. 2000.
[5] I. B. Collings and J. B. Moore, “An HMM approach to adaptive demodu-
lation of QAM signals in fading channels,” Int. J. Adaptive Contr. Signal
Process., vol. 8, pp. 457–474, 1994.
253
[6] , “An adaptive hidden Markov model approach to FM and M-ary
DPSK demodulation in noisy fading channels,” Signal Process., vol. 47,
pp. 71–84, 1995.
[7] A. N. D’Andrea, A. Diglio, and U. Mengali, “Symbol-aided channel
estimation with nonselective Rayleigh fading channels,” IEEE Trans.
Commun., vol. 44, pp. 41–48, Feb. 1995.
[8] A. Doucet, “On sequential simulation-based methods for Bayesian fil-
tering,” Statist. Comput., 2001, to be published.
[9] A. Doucet, J. F. G. de Freitas, and N. Gordon, Sequential Monte Carlo
in Practice. Cambridge, U.K.: Cambridge Univ. Press, 2001.
[10] C. N. Georghiades and J. C. Han, “Sequence estimation in the presence
of random parameters via the EM algorithm,” IEEE Trans. Commun.,
vol. 45, pp. 300–308, Mar. 1997.
[11] M. J. Gertsman and J. H. Lodge, “Symbol-by-symbol MAP demodula-
tion of CPM and PSK signals on Rayleigh flad fading channels,” IEEE
Trans. Commun., vol. 45, pp. 788–799, July 1997.
[12] N. J. Gordon, D. J. Salmon Salmon, and A. F. M. Smith, “A novel ap-
proach to nonlinear/non-Gaussian Bayesian state estimation,” Proc. Inst.
Elect. Eng., Radar Signal Process., vol. 140, pp. 107–113, 1993.
[13] R. Haeb and H. Meyr, “A systematic approach to carrier recovery and de-
tection of digitally phase modulated signals on fading channels,” IEEE
Trans. Commun., vol. 37, pp. 748–754, July 1989.
[14] P. Hoeher and J. H. Lodge, “Turbo DPSK: Iterative differential PSK
demodulation and channel decodings,” IEEE Trans. Commun., vol. 47,
pp. 837–843, June 1999.
[15] M. Hürzeler and H. R. Künsch, “Monte Carlo approximations for gen-
eral state space models,” J. Comput. Graph. Statist., vol. 7, pp. 175–193,
1998.
[16] G. T. Irvine and P. J. McLane, “Symbol-aided plus decision-directed
reception for PSK/TCM modulation on shadowed mobile satellite fading
channels,” IEEE J. Select. Areas Commun., vol. 10, pp. 1289–1299, Oct.
1992.
[17] P. Y. Kam and H. M. Ching, “Sequence estimation over the slow nonse-
lective Rayleigh fading channel with diversity reception and its applica-
tion to Viterbi decoding,” IEEE J. Select. Areas Commun., vol. 10, pp.
562–570, Mar. 1992.
[18] A. Kong, J. S. Liu, and W. H. Wong, “Sequential imputations and
Bayesian missing data problems,” J. Amer. Statist. Assoc., vol. 89, pp.
278–288, 1994.
[19] H. Kong and E. Shwedyk, “On channel estimation and sequence detec-
tion of interleaved coded signals over frequency nonselective Rayleigh
fading channels,” IEEE Trans. Veh. Technol., vol. 47, pp. 558–565, May
1998.
[20] J. S. Liu, Monte Carlo Strategies for Scientific Computing. New York:
Springer-Verlag, 2001.
[21] J. S. Liu, R. Chen, and T. Logvinenko, “A theoretical framework for
sequential importance sampling and resampling,,” in Sequential Monte
Carlo in Practice, A. Doucet, N. Freitas, and N. Gordon, Eds. New
York: Springer-Verlag, 2001, pp. 225–246.
[22] J. S. Liu and R. Chen, “Blind deconvolution via sequential imputations,”
J. Amer. Statist. Assoc., vol. 90, pp. 567–576, 1995.
[23] , “Sequential Monte Carlo methods for dynamic systems,” J. Amer.
Statist. Assoc., vol. 93, pp. 1032–1044, 1998.
[24] Y. Liu and S. D. Bolstein, “Identification of frequency nonselective
fading channels using decision feedback and adaptive linear prediction,”
IEEE Trans. Commun., vol. 43, pp. 1484–1492, Feb./Mar./Apr. 1995.
[25] J. H. Lodge and M. L. Moher, “Maximum likelihood sequence estima-
tion of CPM signals transmitted over Rayleigh flat-fading channels,”
IEEE Trans. Commun., vol. 38, pp. 787–794, June 1990.
[26] D. Makrakis, P. T. Mathiopoulos, and D. P. Bouras, “Optimal decoding
of coded PSK and QAM signals in correlated fast fading channels and
AWGN: A combined envelop, multiple differential and coherent detec-
tion approach,” IEEE Trans. Commun., vol. 42, pp. 63–75, Jan. 1994.
[27] M. L. Moher and J. H. Lodge, “TCMP—A modulation and coding
strategy for Rician fading channels,” IEEE J. Select. Areas Commun.,
vol. 7, pp. 1347–1355, Dec. 1989.
[28] M. K. Pitt and N. Shephard, “Filtering via simulation: Auxiliary particle
filters,” J. Amer. Statist. Assoc., vol. 94, pp. 590–599, 1999.
[29] R. Raheli, A. Polydoros, and C. Tzou, “Per-survivor processing: A
general approach to MLSE in uncertain environments,” IEEE Trans.
Commun., vol. COM-43, pp. 354–364, Feb./Mar./Apr. 1995.
[30] G. M. Vitetta and D. P. Taylor, “Maximum likelihood decoding of un-
coded and coded PSK signal sequences transmitted over Rayleigh flat-
fading channels,” IEEE Trans. Commun., vol. 43, pp. 2750–2758, Nov.
1995.
254
Xiaodong Wang (M’98) received the B.S. degree in
electrical engineering and applied mathematics (with
the highest honor) from Shanghai Jiao Tong Uni-
versity, Shanghai, China, in 1992, the M.S. degree
in electrical engineering from Purdue University,
West Lafayette, IN, in 1995, and the Ph.D. degree
in electrical engineering from Princeton University,
Princeton, NJ, in 1998.
From July 1998 to December 2001, he was an
Assistant Professor with the Department of Elec-
trical Engineering, Texas A&M University, College
Station. In January 2002, he joined the Department of Electrical Engineering,
Columbia University, New York, NY, as an Assistant Professor. His research
interests fall in the general areas of computing, signal processing, and commu-
nications. He has worked in the areas of digital communications, digital signal
processing, parallel and distributed computing, nanoelectronics, and quantum
computing. His current research interests include multiuser communications
theory and advanced signal processing for wireless communications. He wwas
with AT&T Labs—Research, Red Bank, NJ, during the summer of 1997.
Dr. Wang is a member of the American Association for the Advancement
of Science. He received the 1999 NSF CAREER Award and the 2001 IEEE
Communications Society and Information Theory Society Joint Paper Award.
He currently serves as an Associate Editor for the IEEE TRANSACTIONS ON
COMMUNICATIONS and for the IEEE TRANSACTIONS ON SIGNAL PROCESSING.
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 50, NO. 2, FEBRUARY 2002
Rong Chen (M’92) received the B.S. degree in
mathematics from Peking University, Beijing, China,
in 1985 and the M.S. and Ph.D. degrees in statistics
from Carnegie Mellon University, Pittsburgh, PA, in
1987 and 1990, respectively.
He is a Professor at the Department of Informa-
tion and Decision Sciences, College of Business Ad-
ministration, University of Illinois at Chicago (UIC).
Before joining UIC in 1999, he was with the Depart-
ment of Statistics, Texas A&M University, College
Station. His main research interests are in time series
analysis, statistical computing and Monte Carlo methods in dynamic systems,
and statistical applications in engineering and business. He is an Associate Ed-
itor for the Journal of American Statistical Association, the Journal of Business
and Economic Statistics, Statistica Sinica, and Computational Statistics.
Dong Guo received the B.S. degree in geophysics
and computer science (minor) in 1993 from the
Chinese University of Mining and Technology
(CUMT), Xuzhou, China, and the M.S. degree in
geophysics in 1998 from the Graduate School of
Research Institute of Petroleum Exploration and
Development (RIPED), Beijing, China, in 1996. He
received the Ph.D. degree in applied mathematics
from Peking University, Beijing, China, in 1999.
He is pursing a second Ph.D. degree with the
Department of Electrical Engineering, Texas A&M
University, College Station.
His research interests are in the area of statistical signal processing and its
applications, primarily in digital communications.